4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
41 #include <linux/kernel_stat.h>
43 #include <linux/hugetlb.h>
44 #include <linux/mman.h>
45 #include <linux/swap.h>
46 #include <linux/highmem.h>
47 #include <linux/pagemap.h>
48 #include <linux/rmap.h>
49 #include <linux/module.h>
50 #include <linux/init.h>
52 #include <asm/pgalloc.h>
53 #include <asm/uaccess.h>
55 #include <asm/tlbflush.h>
56 #include <asm/pgtable.h>
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr;
66 EXPORT_SYMBOL(max_mapnr);
67 EXPORT_SYMBOL(mem_map);
70 unsigned long num_physpages;
72 * A number of key systems in x86 including ioremap() rely on the assumption
73 * that high_memory defines the upper bound on direct map memory, then end
74 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
75 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
79 unsigned long vmalloc_earlyreserve;
81 EXPORT_SYMBOL(num_physpages);
82 EXPORT_SYMBOL(high_memory);
83 EXPORT_SYMBOL(vmalloc_earlyreserve);
86 * If a p?d_bad entry is found while walking page tables, report
87 * the error, before resetting entry to p?d_none. Usually (but
88 * very seldom) called out from the p?d_none_or_clear_bad macros.
91 void pgd_clear_bad(pgd_t *pgd)
97 void pud_clear_bad(pud_t *pud)
103 void pmd_clear_bad(pmd_t *pmd)
110 * Note: this doesn't free the actual pages themselves. That
111 * has been handled earlier when unmapping all the memory regions.
113 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
115 struct page *page = pmd_page(*pmd);
117 pte_lock_deinit(page);
118 pte_free_tlb(tlb, page);
119 dec_page_state(nr_page_table_pages);
123 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
124 unsigned long addr, unsigned long end,
125 unsigned long floor, unsigned long ceiling)
132 pmd = pmd_offset(pud, addr);
134 next = pmd_addr_end(addr, end);
135 if (pmd_none_or_clear_bad(pmd))
137 free_pte_range(tlb, pmd);
138 } while (pmd++, addr = next, addr != end);
148 if (end - 1 > ceiling - 1)
151 pmd = pmd_offset(pud, start);
153 pmd_free_tlb(tlb, pmd);
156 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
157 unsigned long addr, unsigned long end,
158 unsigned long floor, unsigned long ceiling)
165 pud = pud_offset(pgd, addr);
167 next = pud_addr_end(addr, end);
168 if (pud_none_or_clear_bad(pud))
170 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
171 } while (pud++, addr = next, addr != end);
177 ceiling &= PGDIR_MASK;
181 if (end - 1 > ceiling - 1)
184 pud = pud_offset(pgd, start);
186 pud_free_tlb(tlb, pud);
190 * This function frees user-level page tables of a process.
192 * Must be called with pagetable lock held.
194 void free_pgd_range(struct mmu_gather **tlb,
195 unsigned long addr, unsigned long end,
196 unsigned long floor, unsigned long ceiling)
203 * The next few lines have given us lots of grief...
205 * Why are we testing PMD* at this top level? Because often
206 * there will be no work to do at all, and we'd prefer not to
207 * go all the way down to the bottom just to discover that.
209 * Why all these "- 1"s? Because 0 represents both the bottom
210 * of the address space and the top of it (using -1 for the
211 * top wouldn't help much: the masks would do the wrong thing).
212 * The rule is that addr 0 and floor 0 refer to the bottom of
213 * the address space, but end 0 and ceiling 0 refer to the top
214 * Comparisons need to use "end - 1" and "ceiling - 1" (though
215 * that end 0 case should be mythical).
217 * Wherever addr is brought up or ceiling brought down, we must
218 * be careful to reject "the opposite 0" before it confuses the
219 * subsequent tests. But what about where end is brought down
220 * by PMD_SIZE below? no, end can't go down to 0 there.
222 * Whereas we round start (addr) and ceiling down, by different
223 * masks at different levels, in order to test whether a table
224 * now has no other vmas using it, so can be freed, we don't
225 * bother to round floor or end up - the tests don't need that.
239 if (end - 1 > ceiling - 1)
245 pgd = pgd_offset((*tlb)->mm, addr);
247 next = pgd_addr_end(addr, end);
248 if (pgd_none_or_clear_bad(pgd))
250 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
251 } while (pgd++, addr = next, addr != end);
254 flush_tlb_pgtables((*tlb)->mm, start, end);
257 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
258 unsigned long floor, unsigned long ceiling)
261 struct vm_area_struct *next = vma->vm_next;
262 unsigned long addr = vma->vm_start;
265 * Hide vma from rmap and vmtruncate before freeing pgtables
267 anon_vma_unlink(vma);
268 unlink_file_vma(vma);
270 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
271 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
272 floor, next? next->vm_start: ceiling);
275 * Optimization: gather nearby vmas into one call down
277 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
278 && !is_hugepage_only_range(vma->vm_mm, next->vm_start,
282 anon_vma_unlink(vma);
283 unlink_file_vma(vma);
285 free_pgd_range(tlb, addr, vma->vm_end,
286 floor, next? next->vm_start: ceiling);
292 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
294 struct page *new = pte_alloc_one(mm, address);
299 spin_lock(&mm->page_table_lock);
300 if (pmd_present(*pmd)) { /* Another has populated it */
301 pte_lock_deinit(new);
305 inc_page_state(nr_page_table_pages);
306 pmd_populate(mm, pmd, new);
308 spin_unlock(&mm->page_table_lock);
312 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
314 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
318 spin_lock(&init_mm.page_table_lock);
319 if (pmd_present(*pmd)) /* Another has populated it */
320 pte_free_kernel(new);
322 pmd_populate_kernel(&init_mm, pmd, new);
323 spin_unlock(&init_mm.page_table_lock);
327 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
330 add_mm_counter(mm, file_rss, file_rss);
332 add_mm_counter(mm, anon_rss, anon_rss);
336 * This function is called to print an error when a bad pte
337 * is found. For example, we might have a PFN-mapped pte in
338 * a region that doesn't allow it.
340 * The calling function must still handle the error.
342 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
344 printk(KERN_ERR "Bad pte = %08llx, process = %s, "
345 "vm_flags = %lx, vaddr = %lx\n",
346 (long long)pte_val(pte),
347 (vma->vm_mm == current->mm ? current->comm : "???"),
348 vma->vm_flags, vaddr);
353 * This function gets the "struct page" associated with a pte.
355 * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
356 * will have each page table entry just pointing to a raw page frame
357 * number, and as far as the VM layer is concerned, those do not have
358 * pages associated with them - even if the PFN might point to memory
359 * that otherwise is perfectly fine and has a "struct page".
361 * The way we recognize those mappings is through the rules set up
362 * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
363 * and the vm_pgoff will point to the first PFN mapped: thus every
364 * page that is a raw mapping will always honor the rule
366 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
368 * and if that isn't true, the page has been COW'ed (in which case it
369 * _does_ have a "struct page" associated with it even if it is in a
372 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
374 unsigned long pfn = pte_pfn(pte);
376 if (vma->vm_flags & VM_PFNMAP) {
377 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
378 if (pfn == vma->vm_pgoff + off)
380 if (vma->vm_flags & VM_SHARED)
385 * Add some anal sanity checks for now. Eventually,
386 * we should just do "return pfn_to_page(pfn)", but
387 * in the meantime we check that we get a valid pfn,
388 * and that the resulting page looks ok.
390 * Remove this test eventually!
392 if (unlikely(!pfn_valid(pfn))) {
393 print_bad_pte(vma, pte, addr);
398 * NOTE! We still have PageReserved() pages in the page
401 * The PAGE_ZERO() pages and various VDSO mappings can
402 * cause them to exist.
404 return pfn_to_page(pfn);
408 * copy one vm_area from one task to the other. Assumes the page tables
409 * already present in the new task to be cleared in the whole range
410 * covered by this vma.
414 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
415 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
416 unsigned long addr, int *rss)
418 unsigned long vm_flags = vma->vm_flags;
419 pte_t pte = *src_pte;
422 /* pte contains position in swap or file, so copy. */
423 if (unlikely(!pte_present(pte))) {
424 if (!pte_file(pte)) {
425 swap_duplicate(pte_to_swp_entry(pte));
426 /* make sure dst_mm is on swapoff's mmlist. */
427 if (unlikely(list_empty(&dst_mm->mmlist))) {
428 spin_lock(&mmlist_lock);
429 if (list_empty(&dst_mm->mmlist))
430 list_add(&dst_mm->mmlist,
432 spin_unlock(&mmlist_lock);
439 * If it's a COW mapping, write protect it both
440 * in the parent and the child
442 if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
443 ptep_set_wrprotect(src_mm, addr, src_pte);
448 * If it's a shared mapping, mark it clean in
451 if (vm_flags & VM_SHARED)
452 pte = pte_mkclean(pte);
453 pte = pte_mkold(pte);
455 page = vm_normal_page(vma, addr, pte);
459 rss[!!PageAnon(page)]++;
463 set_pte_at(dst_mm, addr, dst_pte, pte);
466 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
467 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
468 unsigned long addr, unsigned long end)
470 pte_t *src_pte, *dst_pte;
471 spinlock_t *src_ptl, *dst_ptl;
477 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
480 src_pte = pte_offset_map_nested(src_pmd, addr);
481 src_ptl = pte_lockptr(src_mm, src_pmd);
486 * We are holding two locks at this point - either of them
487 * could generate latencies in another task on another CPU.
489 if (progress >= 32) {
491 if (need_resched() ||
492 need_lockbreak(src_ptl) ||
493 need_lockbreak(dst_ptl))
496 if (pte_none(*src_pte)) {
500 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
502 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
504 spin_unlock(src_ptl);
505 pte_unmap_nested(src_pte - 1);
506 add_mm_rss(dst_mm, rss[0], rss[1]);
507 pte_unmap_unlock(dst_pte - 1, dst_ptl);
514 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
515 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
516 unsigned long addr, unsigned long end)
518 pmd_t *src_pmd, *dst_pmd;
521 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
524 src_pmd = pmd_offset(src_pud, addr);
526 next = pmd_addr_end(addr, end);
527 if (pmd_none_or_clear_bad(src_pmd))
529 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
532 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
536 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
537 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
538 unsigned long addr, unsigned long end)
540 pud_t *src_pud, *dst_pud;
543 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
546 src_pud = pud_offset(src_pgd, addr);
548 next = pud_addr_end(addr, end);
549 if (pud_none_or_clear_bad(src_pud))
551 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
554 } while (dst_pud++, src_pud++, addr = next, addr != end);
558 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
559 struct vm_area_struct *vma)
561 pgd_t *src_pgd, *dst_pgd;
563 unsigned long addr = vma->vm_start;
564 unsigned long end = vma->vm_end;
567 * Don't copy ptes where a page fault will fill them correctly.
568 * Fork becomes much lighter when there are big shared or private
569 * readonly mappings. The tradeoff is that copy_page_range is more
570 * efficient than faulting.
572 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP))) {
577 if (is_vm_hugetlb_page(vma))
578 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
580 dst_pgd = pgd_offset(dst_mm, addr);
581 src_pgd = pgd_offset(src_mm, addr);
583 next = pgd_addr_end(addr, end);
584 if (pgd_none_or_clear_bad(src_pgd))
586 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
589 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
593 static unsigned long zap_pte_range(struct mmu_gather *tlb,
594 struct vm_area_struct *vma, pmd_t *pmd,
595 unsigned long addr, unsigned long end,
596 long *zap_work, struct zap_details *details)
598 struct mm_struct *mm = tlb->mm;
604 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
607 if (pte_none(ptent)) {
611 if (pte_present(ptent)) {
614 (*zap_work) -= PAGE_SIZE;
616 page = vm_normal_page(vma, addr, ptent);
617 if (unlikely(details) && page) {
619 * unmap_shared_mapping_pages() wants to
620 * invalidate cache without truncating:
621 * unmap shared but keep private pages.
623 if (details->check_mapping &&
624 details->check_mapping != page->mapping)
627 * Each page->index must be checked when
628 * invalidating or truncating nonlinear.
630 if (details->nonlinear_vma &&
631 (page->index < details->first_index ||
632 page->index > details->last_index))
635 ptent = ptep_get_and_clear_full(mm, addr, pte,
637 tlb_remove_tlb_entry(tlb, pte, addr);
640 if (unlikely(details) && details->nonlinear_vma
641 && linear_page_index(details->nonlinear_vma,
642 addr) != page->index)
643 set_pte_at(mm, addr, pte,
644 pgoff_to_pte(page->index));
648 if (pte_dirty(ptent))
649 set_page_dirty(page);
650 if (pte_young(ptent))
651 mark_page_accessed(page);
654 page_remove_rmap(page);
655 tlb_remove_page(tlb, page);
659 * If details->check_mapping, we leave swap entries;
660 * if details->nonlinear_vma, we leave file entries.
662 if (unlikely(details))
664 if (!pte_file(ptent))
665 free_swap_and_cache(pte_to_swp_entry(ptent));
666 pte_clear_full(mm, addr, pte, tlb->fullmm);
667 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
669 add_mm_rss(mm, file_rss, anon_rss);
670 pte_unmap_unlock(pte - 1, ptl);
675 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
676 struct vm_area_struct *vma, pud_t *pud,
677 unsigned long addr, unsigned long end,
678 long *zap_work, struct zap_details *details)
683 pmd = pmd_offset(pud, addr);
685 next = pmd_addr_end(addr, end);
686 if (pmd_none_or_clear_bad(pmd)) {
690 next = zap_pte_range(tlb, vma, pmd, addr, next,
692 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
697 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
698 struct vm_area_struct *vma, pgd_t *pgd,
699 unsigned long addr, unsigned long end,
700 long *zap_work, struct zap_details *details)
705 pud = pud_offset(pgd, addr);
707 next = pud_addr_end(addr, end);
708 if (pud_none_or_clear_bad(pud)) {
712 next = zap_pmd_range(tlb, vma, pud, addr, next,
714 } while (pud++, addr = next, (addr != end && *zap_work > 0));
719 static unsigned long unmap_page_range(struct mmu_gather *tlb,
720 struct vm_area_struct *vma,
721 unsigned long addr, unsigned long end,
722 long *zap_work, struct zap_details *details)
727 if (details && !details->check_mapping && !details->nonlinear_vma)
731 tlb_start_vma(tlb, vma);
732 pgd = pgd_offset(vma->vm_mm, addr);
734 next = pgd_addr_end(addr, end);
735 if (pgd_none_or_clear_bad(pgd)) {
739 next = zap_pud_range(tlb, vma, pgd, addr, next,
741 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
742 tlb_end_vma(tlb, vma);
747 #ifdef CONFIG_PREEMPT
748 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
750 /* No preempt: go for improved straight-line efficiency */
751 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
755 * unmap_vmas - unmap a range of memory covered by a list of vma's
756 * @tlbp: address of the caller's struct mmu_gather
757 * @vma: the starting vma
758 * @start_addr: virtual address at which to start unmapping
759 * @end_addr: virtual address at which to end unmapping
760 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
761 * @details: details of nonlinear truncation or shared cache invalidation
763 * Returns the end address of the unmapping (restart addr if interrupted).
765 * Unmap all pages in the vma list.
767 * We aim to not hold locks for too long (for scheduling latency reasons).
768 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
769 * return the ending mmu_gather to the caller.
771 * Only addresses between `start' and `end' will be unmapped.
773 * The VMA list must be sorted in ascending virtual address order.
775 * unmap_vmas() assumes that the caller will flush the whole unmapped address
776 * range after unmap_vmas() returns. So the only responsibility here is to
777 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
778 * drops the lock and schedules.
780 unsigned long unmap_vmas(struct mmu_gather **tlbp,
781 struct vm_area_struct *vma, unsigned long start_addr,
782 unsigned long end_addr, unsigned long *nr_accounted,
783 struct zap_details *details)
785 long zap_work = ZAP_BLOCK_SIZE;
786 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
787 int tlb_start_valid = 0;
788 unsigned long start = start_addr;
789 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
790 int fullmm = (*tlbp)->fullmm;
792 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
795 start = max(vma->vm_start, start_addr);
796 if (start >= vma->vm_end)
798 end = min(vma->vm_end, end_addr);
799 if (end <= vma->vm_start)
802 if (vma->vm_flags & VM_ACCOUNT)
803 *nr_accounted += (end - start) >> PAGE_SHIFT;
805 while (start != end) {
806 if (!tlb_start_valid) {
811 if (unlikely(is_vm_hugetlb_page(vma))) {
812 unmap_hugepage_range(vma, start, end);
813 zap_work -= (end - start) /
814 (HPAGE_SIZE / PAGE_SIZE);
817 start = unmap_page_range(*tlbp, vma,
818 start, end, &zap_work, details);
821 BUG_ON(start != end);
825 tlb_finish_mmu(*tlbp, tlb_start, start);
827 if (need_resched() ||
828 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
836 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
838 zap_work = ZAP_BLOCK_SIZE;
842 return start; /* which is now the end (or restart) address */
846 * zap_page_range - remove user pages in a given range
847 * @vma: vm_area_struct holding the applicable pages
848 * @address: starting address of pages to zap
849 * @size: number of bytes to zap
850 * @details: details of nonlinear truncation or shared cache invalidation
852 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
853 unsigned long size, struct zap_details *details)
855 struct mm_struct *mm = vma->vm_mm;
856 struct mmu_gather *tlb;
857 unsigned long end = address + size;
858 unsigned long nr_accounted = 0;
861 tlb = tlb_gather_mmu(mm, 0);
862 update_hiwater_rss(mm);
863 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
865 tlb_finish_mmu(tlb, address, end);
870 * Do a quick page-table lookup for a single page.
872 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
881 struct mm_struct *mm = vma->vm_mm;
883 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
885 BUG_ON(flags & FOLL_GET);
890 pgd = pgd_offset(mm, address);
891 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
894 pud = pud_offset(pgd, address);
895 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
898 pmd = pmd_offset(pud, address);
899 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
902 if (pmd_huge(*pmd)) {
903 BUG_ON(flags & FOLL_GET);
904 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
908 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
913 if (!pte_present(pte))
915 if ((flags & FOLL_WRITE) && !pte_write(pte))
917 page = vm_normal_page(vma, address, pte);
921 if (flags & FOLL_GET)
923 if (flags & FOLL_TOUCH) {
924 if ((flags & FOLL_WRITE) &&
925 !pte_dirty(pte) && !PageDirty(page))
926 set_page_dirty(page);
927 mark_page_accessed(page);
930 pte_unmap_unlock(ptep, ptl);
936 * When core dumping an enormous anonymous area that nobody
937 * has touched so far, we don't want to allocate page tables.
939 if (flags & FOLL_ANON) {
940 page = ZERO_PAGE(address);
941 if (flags & FOLL_GET)
943 BUG_ON(flags & FOLL_WRITE);
948 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
949 unsigned long start, int len, int write, int force,
950 struct page **pages, struct vm_area_struct **vmas)
953 unsigned int vm_flags;
956 * Require read or write permissions.
957 * If 'force' is set, we only require the "MAY" flags.
959 vm_flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
960 vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
964 struct vm_area_struct *vma;
965 unsigned int foll_flags;
967 vma = find_extend_vma(mm, start);
968 if (!vma && in_gate_area(tsk, start)) {
969 unsigned long pg = start & PAGE_MASK;
970 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
975 if (write) /* user gate pages are read-only */
976 return i ? : -EFAULT;
978 pgd = pgd_offset_k(pg);
980 pgd = pgd_offset_gate(mm, pg);
981 BUG_ON(pgd_none(*pgd));
982 pud = pud_offset(pgd, pg);
983 BUG_ON(pud_none(*pud));
984 pmd = pmd_offset(pud, pg);
986 return i ? : -EFAULT;
987 pte = pte_offset_map(pmd, pg);
988 if (pte_none(*pte)) {
990 return i ? : -EFAULT;
993 struct page *page = vm_normal_page(gate_vma, start, *pte);
1007 if (!vma || (vma->vm_flags & VM_IO)
1008 || !(vm_flags & vma->vm_flags))
1009 return i ? : -EFAULT;
1011 if (is_vm_hugetlb_page(vma)) {
1012 i = follow_hugetlb_page(mm, vma, pages, vmas,
1017 foll_flags = FOLL_TOUCH;
1019 foll_flags |= FOLL_GET;
1020 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1021 (!vma->vm_ops || !vma->vm_ops->nopage))
1022 foll_flags |= FOLL_ANON;
1028 foll_flags |= FOLL_WRITE;
1031 while (!(page = follow_page(vma, start, foll_flags))) {
1033 ret = __handle_mm_fault(mm, vma, start,
1034 foll_flags & FOLL_WRITE);
1036 * The VM_FAULT_WRITE bit tells us that do_wp_page has
1037 * broken COW when necessary, even if maybe_mkwrite
1038 * decided not to set pte_write. We can thus safely do
1039 * subsequent page lookups as if they were reads.
1041 if (ret & VM_FAULT_WRITE)
1042 foll_flags &= ~FOLL_WRITE;
1044 switch (ret & ~VM_FAULT_WRITE) {
1045 case VM_FAULT_MINOR:
1048 case VM_FAULT_MAJOR:
1051 case VM_FAULT_SIGBUS:
1052 return i ? i : -EFAULT;
1054 return i ? i : -ENOMEM;
1061 flush_dcache_page(page);
1068 } while (len && start < vma->vm_end);
1072 EXPORT_SYMBOL(get_user_pages);
1074 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1075 unsigned long addr, unsigned long end, pgprot_t prot)
1080 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1084 struct page *page = ZERO_PAGE(addr);
1085 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1086 page_cache_get(page);
1087 page_add_file_rmap(page);
1088 inc_mm_counter(mm, file_rss);
1089 BUG_ON(!pte_none(*pte));
1090 set_pte_at(mm, addr, pte, zero_pte);
1091 } while (pte++, addr += PAGE_SIZE, addr != end);
1092 pte_unmap_unlock(pte - 1, ptl);
1096 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1097 unsigned long addr, unsigned long end, pgprot_t prot)
1102 pmd = pmd_alloc(mm, pud, addr);
1106 next = pmd_addr_end(addr, end);
1107 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1109 } while (pmd++, addr = next, addr != end);
1113 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1114 unsigned long addr, unsigned long end, pgprot_t prot)
1119 pud = pud_alloc(mm, pgd, addr);
1123 next = pud_addr_end(addr, end);
1124 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1126 } while (pud++, addr = next, addr != end);
1130 int zeromap_page_range(struct vm_area_struct *vma,
1131 unsigned long addr, unsigned long size, pgprot_t prot)
1135 unsigned long end = addr + size;
1136 struct mm_struct *mm = vma->vm_mm;
1139 BUG_ON(addr >= end);
1140 pgd = pgd_offset(mm, addr);
1141 flush_cache_range(vma, addr, end);
1143 next = pgd_addr_end(addr, end);
1144 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1147 } while (pgd++, addr = next, addr != end);
1151 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1153 pgd_t * pgd = pgd_offset(mm, addr);
1154 pud_t * pud = pud_alloc(mm, pgd, addr);
1156 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1158 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1164 * This is the old fallback for page remapping.
1166 * For historical reasons, it only allows reserved pages. Only
1167 * old drivers should use this, and they needed to mark their
1168 * pages reserved for the old functions anyway.
1170 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1180 flush_dcache_page(page);
1181 pte = get_locked_pte(mm, addr, &ptl);
1185 if (!pte_none(*pte))
1188 /* Ok, finally just insert the thing.. */
1190 inc_mm_counter(mm, file_rss);
1191 page_add_file_rmap(page);
1192 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1196 pte_unmap_unlock(pte, ptl);
1202 * This allows drivers to insert individual pages they've allocated
1205 * The page has to be a nice clean _individual_ kernel allocation.
1206 * If you allocate a compound page, you need to have marked it as
1207 * such (__GFP_COMP), or manually just split the page up yourself
1208 * (which is mainly an issue of doing "set_page_count(page, 1)" for
1209 * each sub-page, and then freeing them one by one when you free
1210 * them rather than freeing it as a compound page).
1212 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1213 * took an arbitrary page protection parameter. This doesn't allow
1214 * that. Your vma protection will have to be set up correctly, which
1215 * means that if you want a shared writable mapping, you'd better
1216 * ask for a shared writable mapping!
1218 * The page does not need to be reserved.
1220 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1222 if (addr < vma->vm_start || addr >= vma->vm_end)
1224 if (!page_count(page))
1226 return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1228 EXPORT_SYMBOL(vm_insert_page);
1231 * Somebody does a pfn remapping that doesn't actually work as a vma.
1233 * Do it as individual pages instead, and warn about it. It's bad form,
1234 * and very inefficient.
1236 static int incomplete_pfn_remap(struct vm_area_struct *vma,
1237 unsigned long start, unsigned long end,
1238 unsigned long pfn, pgprot_t prot)
1240 static int warn = 10;
1244 if (!(vma->vm_flags & VM_INCOMPLETE)) {
1247 printk("%s does an incomplete pfn remapping", current->comm);
1251 vma->vm_flags |= VM_INCOMPLETE | VM_IO | VM_RESERVED;
1253 if (start < vma->vm_start || end > vma->vm_end)
1256 if (!pfn_valid(pfn))
1259 page = pfn_to_page(pfn);
1260 if (!PageReserved(page))
1264 while (start < end) {
1265 retval = insert_page(vma->vm_mm, start, page, prot);
1275 * maps a range of physical memory into the requested pages. the old
1276 * mappings are removed. any references to nonexistent pages results
1277 * in null mappings (currently treated as "copy-on-access")
1279 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1280 unsigned long addr, unsigned long end,
1281 unsigned long pfn, pgprot_t prot)
1286 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1290 BUG_ON(!pte_none(*pte));
1291 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1293 } while (pte++, addr += PAGE_SIZE, addr != end);
1294 pte_unmap_unlock(pte - 1, ptl);
1298 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1299 unsigned long addr, unsigned long end,
1300 unsigned long pfn, pgprot_t prot)
1305 pfn -= addr >> PAGE_SHIFT;
1306 pmd = pmd_alloc(mm, pud, addr);
1310 next = pmd_addr_end(addr, end);
1311 if (remap_pte_range(mm, pmd, addr, next,
1312 pfn + (addr >> PAGE_SHIFT), prot))
1314 } while (pmd++, addr = next, addr != end);
1318 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1319 unsigned long addr, unsigned long end,
1320 unsigned long pfn, pgprot_t prot)
1325 pfn -= addr >> PAGE_SHIFT;
1326 pud = pud_alloc(mm, pgd, addr);
1330 next = pud_addr_end(addr, end);
1331 if (remap_pmd_range(mm, pud, addr, next,
1332 pfn + (addr >> PAGE_SHIFT), prot))
1334 } while (pud++, addr = next, addr != end);
1338 /* Note: this is only safe if the mm semaphore is held when called. */
1339 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1340 unsigned long pfn, unsigned long size, pgprot_t prot)
1344 unsigned long end = addr + PAGE_ALIGN(size);
1345 struct mm_struct *mm = vma->vm_mm;
1349 * Physically remapped pages are special. Tell the
1350 * rest of the world about it:
1351 * VM_IO tells people not to look at these pages
1352 * (accesses can have side effects).
1353 * VM_RESERVED is specified all over the place, because
1354 * in 2.4 it kept swapout's vma scan off this vma; but
1355 * in 2.6 the LRU scan won't even find its pages, so this
1356 * flag means no more than count its pages in reserved_vm,
1357 * and omit it from core dump, even when VM_IO turned off.
1358 * VM_PFNMAP tells the core MM that the base pages are just
1359 * raw PFN mappings, and do not have a "struct page" associated
1362 * There's a horrible special case to handle copy-on-write
1363 * behaviour that some programs depend on. We mark the "original"
1364 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1366 if (!(vma->vm_flags & VM_SHARED)) {
1367 if (addr != vma->vm_start || end != vma->vm_end)
1368 return incomplete_pfn_remap(vma, addr, end, pfn, prot);
1369 vma->vm_pgoff = pfn;
1372 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1374 BUG_ON(addr >= end);
1375 pfn -= addr >> PAGE_SHIFT;
1376 pgd = pgd_offset(mm, addr);
1377 flush_cache_range(vma, addr, end);
1379 next = pgd_addr_end(addr, end);
1380 err = remap_pud_range(mm, pgd, addr, next,
1381 pfn + (addr >> PAGE_SHIFT), prot);
1384 } while (pgd++, addr = next, addr != end);
1387 EXPORT_SYMBOL(remap_pfn_range);
1390 * handle_pte_fault chooses page fault handler according to an entry
1391 * which was read non-atomically. Before making any commitment, on
1392 * those architectures or configurations (e.g. i386 with PAE) which
1393 * might give a mix of unmatched parts, do_swap_page and do_file_page
1394 * must check under lock before unmapping the pte and proceeding
1395 * (but do_wp_page is only called after already making such a check;
1396 * and do_anonymous_page and do_no_page can safely check later on).
1398 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1399 pte_t *page_table, pte_t orig_pte)
1402 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1403 if (sizeof(pte_t) > sizeof(unsigned long)) {
1404 spinlock_t *ptl = pte_lockptr(mm, pmd);
1406 same = pte_same(*page_table, orig_pte);
1410 pte_unmap(page_table);
1415 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1416 * servicing faults for write access. In the normal case, do always want
1417 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1418 * that do not have writing enabled, when used by access_process_vm.
1420 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1422 if (likely(vma->vm_flags & VM_WRITE))
1423 pte = pte_mkwrite(pte);
1427 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1430 * If the source page was a PFN mapping, we don't have
1431 * a "struct page" for it. We do a best-effort copy by
1432 * just copying from the original user address. If that
1433 * fails, we just zero-fill it. Live with it.
1435 if (unlikely(!src)) {
1436 void *kaddr = kmap_atomic(dst, KM_USER0);
1437 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1440 * This really shouldn't fail, because the page is there
1441 * in the page tables. But it might just be unreadable,
1442 * in which case we just give up and fill the result with
1445 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1446 memset(kaddr, 0, PAGE_SIZE);
1447 kunmap_atomic(kaddr, KM_USER0);
1451 copy_user_highpage(dst, src, va);
1455 * This routine handles present pages, when users try to write
1456 * to a shared page. It is done by copying the page to a new address
1457 * and decrementing the shared-page counter for the old page.
1459 * Note that this routine assumes that the protection checks have been
1460 * done by the caller (the low-level page fault routine in most cases).
1461 * Thus we can safely just mark it writable once we've done any necessary
1464 * We also mark the page dirty at this point even though the page will
1465 * change only once the write actually happens. This avoids a few races,
1466 * and potentially makes it more efficient.
1468 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1469 * but allow concurrent faults), with pte both mapped and locked.
1470 * We return with mmap_sem still held, but pte unmapped and unlocked.
1472 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1473 unsigned long address, pte_t *page_table, pmd_t *pmd,
1474 spinlock_t *ptl, pte_t orig_pte)
1476 struct page *old_page, *new_page;
1478 int ret = VM_FAULT_MINOR;
1480 old_page = vm_normal_page(vma, address, orig_pte);
1484 if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1485 int reuse = can_share_swap_page(old_page);
1486 unlock_page(old_page);
1488 flush_cache_page(vma, address, pte_pfn(orig_pte));
1489 entry = pte_mkyoung(orig_pte);
1490 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1491 ptep_set_access_flags(vma, address, page_table, entry, 1);
1492 update_mmu_cache(vma, address, entry);
1493 lazy_mmu_prot_update(entry);
1494 ret |= VM_FAULT_WRITE;
1500 * Ok, we need to copy. Oh, well..
1502 page_cache_get(old_page);
1504 pte_unmap_unlock(page_table, ptl);
1506 if (unlikely(anon_vma_prepare(vma)))
1508 if (old_page == ZERO_PAGE(address)) {
1509 new_page = alloc_zeroed_user_highpage(vma, address);
1513 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1516 cow_user_page(new_page, old_page, address);
1520 * Re-check the pte - we dropped the lock
1522 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1523 if (likely(pte_same(*page_table, orig_pte))) {
1525 page_remove_rmap(old_page);
1526 if (!PageAnon(old_page)) {
1527 dec_mm_counter(mm, file_rss);
1528 inc_mm_counter(mm, anon_rss);
1531 inc_mm_counter(mm, anon_rss);
1532 flush_cache_page(vma, address, pte_pfn(orig_pte));
1533 entry = mk_pte(new_page, vma->vm_page_prot);
1534 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1535 ptep_establish(vma, address, page_table, entry);
1536 update_mmu_cache(vma, address, entry);
1537 lazy_mmu_prot_update(entry);
1538 lru_cache_add_active(new_page);
1539 page_add_anon_rmap(new_page, vma, address);
1541 /* Free the old page.. */
1542 new_page = old_page;
1543 ret |= VM_FAULT_WRITE;
1546 page_cache_release(new_page);
1548 page_cache_release(old_page);
1550 pte_unmap_unlock(page_table, ptl);
1554 page_cache_release(old_page);
1555 return VM_FAULT_OOM;
1559 * Helper functions for unmap_mapping_range().
1561 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1563 * We have to restart searching the prio_tree whenever we drop the lock,
1564 * since the iterator is only valid while the lock is held, and anyway
1565 * a later vma might be split and reinserted earlier while lock dropped.
1567 * The list of nonlinear vmas could be handled more efficiently, using
1568 * a placeholder, but handle it in the same way until a need is shown.
1569 * It is important to search the prio_tree before nonlinear list: a vma
1570 * may become nonlinear and be shifted from prio_tree to nonlinear list
1571 * while the lock is dropped; but never shifted from list to prio_tree.
1573 * In order to make forward progress despite restarting the search,
1574 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1575 * quickly skip it next time around. Since the prio_tree search only
1576 * shows us those vmas affected by unmapping the range in question, we
1577 * can't efficiently keep all vmas in step with mapping->truncate_count:
1578 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1579 * mapping->truncate_count and vma->vm_truncate_count are protected by
1582 * In order to make forward progress despite repeatedly restarting some
1583 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1584 * and restart from that address when we reach that vma again. It might
1585 * have been split or merged, shrunk or extended, but never shifted: so
1586 * restart_addr remains valid so long as it remains in the vma's range.
1587 * unmap_mapping_range forces truncate_count to leap over page-aligned
1588 * values so we can save vma's restart_addr in its truncate_count field.
1590 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1592 static void reset_vma_truncate_counts(struct address_space *mapping)
1594 struct vm_area_struct *vma;
1595 struct prio_tree_iter iter;
1597 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1598 vma->vm_truncate_count = 0;
1599 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1600 vma->vm_truncate_count = 0;
1603 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1604 unsigned long start_addr, unsigned long end_addr,
1605 struct zap_details *details)
1607 unsigned long restart_addr;
1611 restart_addr = vma->vm_truncate_count;
1612 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1613 start_addr = restart_addr;
1614 if (start_addr >= end_addr) {
1615 /* Top of vma has been split off since last time */
1616 vma->vm_truncate_count = details->truncate_count;
1621 restart_addr = zap_page_range(vma, start_addr,
1622 end_addr - start_addr, details);
1623 need_break = need_resched() ||
1624 need_lockbreak(details->i_mmap_lock);
1626 if (restart_addr >= end_addr) {
1627 /* We have now completed this vma: mark it so */
1628 vma->vm_truncate_count = details->truncate_count;
1632 /* Note restart_addr in vma's truncate_count field */
1633 vma->vm_truncate_count = restart_addr;
1638 spin_unlock(details->i_mmap_lock);
1640 spin_lock(details->i_mmap_lock);
1644 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1645 struct zap_details *details)
1647 struct vm_area_struct *vma;
1648 struct prio_tree_iter iter;
1649 pgoff_t vba, vea, zba, zea;
1652 vma_prio_tree_foreach(vma, &iter, root,
1653 details->first_index, details->last_index) {
1654 /* Skip quickly over those we have already dealt with */
1655 if (vma->vm_truncate_count == details->truncate_count)
1658 vba = vma->vm_pgoff;
1659 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1660 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1661 zba = details->first_index;
1664 zea = details->last_index;
1668 if (unmap_mapping_range_vma(vma,
1669 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1670 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1676 static inline void unmap_mapping_range_list(struct list_head *head,
1677 struct zap_details *details)
1679 struct vm_area_struct *vma;
1682 * In nonlinear VMAs there is no correspondence between virtual address
1683 * offset and file offset. So we must perform an exhaustive search
1684 * across *all* the pages in each nonlinear VMA, not just the pages
1685 * whose virtual address lies outside the file truncation point.
1688 list_for_each_entry(vma, head, shared.vm_set.list) {
1689 /* Skip quickly over those we have already dealt with */
1690 if (vma->vm_truncate_count == details->truncate_count)
1692 details->nonlinear_vma = vma;
1693 if (unmap_mapping_range_vma(vma, vma->vm_start,
1694 vma->vm_end, details) < 0)
1700 * unmap_mapping_range - unmap the portion of all mmaps
1701 * in the specified address_space corresponding to the specified
1702 * page range in the underlying file.
1703 * @mapping: the address space containing mmaps to be unmapped.
1704 * @holebegin: byte in first page to unmap, relative to the start of
1705 * the underlying file. This will be rounded down to a PAGE_SIZE
1706 * boundary. Note that this is different from vmtruncate(), which
1707 * must keep the partial page. In contrast, we must get rid of
1709 * @holelen: size of prospective hole in bytes. This will be rounded
1710 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1712 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1713 * but 0 when invalidating pagecache, don't throw away private data.
1715 void unmap_mapping_range(struct address_space *mapping,
1716 loff_t const holebegin, loff_t const holelen, int even_cows)
1718 struct zap_details details;
1719 pgoff_t hba = holebegin >> PAGE_SHIFT;
1720 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1722 /* Check for overflow. */
1723 if (sizeof(holelen) > sizeof(hlen)) {
1725 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1726 if (holeend & ~(long long)ULONG_MAX)
1727 hlen = ULONG_MAX - hba + 1;
1730 details.check_mapping = even_cows? NULL: mapping;
1731 details.nonlinear_vma = NULL;
1732 details.first_index = hba;
1733 details.last_index = hba + hlen - 1;
1734 if (details.last_index < details.first_index)
1735 details.last_index = ULONG_MAX;
1736 details.i_mmap_lock = &mapping->i_mmap_lock;
1738 spin_lock(&mapping->i_mmap_lock);
1740 /* serialize i_size write against truncate_count write */
1742 /* Protect against page faults, and endless unmapping loops */
1743 mapping->truncate_count++;
1745 * For archs where spin_lock has inclusive semantics like ia64
1746 * this smp_mb() will prevent to read pagetable contents
1747 * before the truncate_count increment is visible to
1751 if (unlikely(is_restart_addr(mapping->truncate_count))) {
1752 if (mapping->truncate_count == 0)
1753 reset_vma_truncate_counts(mapping);
1754 mapping->truncate_count++;
1756 details.truncate_count = mapping->truncate_count;
1758 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1759 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1760 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1761 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1762 spin_unlock(&mapping->i_mmap_lock);
1764 EXPORT_SYMBOL(unmap_mapping_range);
1767 * Handle all mappings that got truncated by a "truncate()"
1770 * NOTE! We have to be ready to update the memory sharing
1771 * between the file and the memory map for a potential last
1772 * incomplete page. Ugly, but necessary.
1774 int vmtruncate(struct inode * inode, loff_t offset)
1776 struct address_space *mapping = inode->i_mapping;
1777 unsigned long limit;
1779 if (inode->i_size < offset)
1782 * truncation of in-use swapfiles is disallowed - it would cause
1783 * subsequent swapout to scribble on the now-freed blocks.
1785 if (IS_SWAPFILE(inode))
1787 i_size_write(inode, offset);
1788 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1789 truncate_inode_pages(mapping, offset);
1793 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1794 if (limit != RLIM_INFINITY && offset > limit)
1796 if (offset > inode->i_sb->s_maxbytes)
1798 i_size_write(inode, offset);
1801 if (inode->i_op && inode->i_op->truncate)
1802 inode->i_op->truncate(inode);
1805 send_sig(SIGXFSZ, current, 0);
1812 EXPORT_SYMBOL(vmtruncate);
1815 * Primitive swap readahead code. We simply read an aligned block of
1816 * (1 << page_cluster) entries in the swap area. This method is chosen
1817 * because it doesn't cost us any seek time. We also make sure to queue
1818 * the 'original' request together with the readahead ones...
1820 * This has been extended to use the NUMA policies from the mm triggering
1823 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1825 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1828 struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1831 struct page *new_page;
1832 unsigned long offset;
1835 * Get the number of handles we should do readahead io to.
1837 num = valid_swaphandles(entry, &offset);
1838 for (i = 0; i < num; offset++, i++) {
1839 /* Ok, do the async read-ahead now */
1840 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1841 offset), vma, addr);
1844 page_cache_release(new_page);
1847 * Find the next applicable VMA for the NUMA policy.
1853 if (addr >= vma->vm_end) {
1855 next_vma = vma ? vma->vm_next : NULL;
1857 if (vma && addr < vma->vm_start)
1860 if (next_vma && addr >= next_vma->vm_start) {
1862 next_vma = vma->vm_next;
1867 lru_add_drain(); /* Push any new pages onto the LRU now */
1871 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1872 * but allow concurrent faults), and pte mapped but not yet locked.
1873 * We return with mmap_sem still held, but pte unmapped and unlocked.
1875 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1876 unsigned long address, pte_t *page_table, pmd_t *pmd,
1877 int write_access, pte_t orig_pte)
1883 int ret = VM_FAULT_MINOR;
1885 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1888 entry = pte_to_swp_entry(orig_pte);
1889 page = lookup_swap_cache(entry);
1891 swapin_readahead(entry, address, vma);
1892 page = read_swap_cache_async(entry, vma, address);
1895 * Back out if somebody else faulted in this pte
1896 * while we released the pte lock.
1898 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1899 if (likely(pte_same(*page_table, orig_pte)))
1904 /* Had to read the page from swap area: Major fault */
1905 ret = VM_FAULT_MAJOR;
1906 inc_page_state(pgmajfault);
1910 mark_page_accessed(page);
1914 * Back out if somebody else already faulted in this pte.
1916 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1917 if (unlikely(!pte_same(*page_table, orig_pte)))
1920 if (unlikely(!PageUptodate(page))) {
1921 ret = VM_FAULT_SIGBUS;
1925 /* The page isn't present yet, go ahead with the fault. */
1927 inc_mm_counter(mm, anon_rss);
1928 pte = mk_pte(page, vma->vm_page_prot);
1929 if (write_access && can_share_swap_page(page)) {
1930 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1934 flush_icache_page(vma, page);
1935 set_pte_at(mm, address, page_table, pte);
1936 page_add_anon_rmap(page, vma, address);
1940 remove_exclusive_swap_page(page);
1944 if (do_wp_page(mm, vma, address,
1945 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1950 /* No need to invalidate - it was non-present before */
1951 update_mmu_cache(vma, address, pte);
1952 lazy_mmu_prot_update(pte);
1954 pte_unmap_unlock(page_table, ptl);
1958 pte_unmap_unlock(page_table, ptl);
1960 page_cache_release(page);
1965 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1966 * but allow concurrent faults), and pte mapped but not yet locked.
1967 * We return with mmap_sem still held, but pte unmapped and unlocked.
1969 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1970 unsigned long address, pte_t *page_table, pmd_t *pmd,
1978 /* Allocate our own private page. */
1979 pte_unmap(page_table);
1981 if (unlikely(anon_vma_prepare(vma)))
1983 page = alloc_zeroed_user_highpage(vma, address);
1987 entry = mk_pte(page, vma->vm_page_prot);
1988 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1990 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1991 if (!pte_none(*page_table))
1993 inc_mm_counter(mm, anon_rss);
1994 lru_cache_add_active(page);
1995 SetPageReferenced(page);
1996 page_add_anon_rmap(page, vma, address);
1998 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1999 page = ZERO_PAGE(address);
2000 page_cache_get(page);
2001 entry = mk_pte(page, vma->vm_page_prot);
2003 ptl = pte_lockptr(mm, pmd);
2005 if (!pte_none(*page_table))
2007 inc_mm_counter(mm, file_rss);
2008 page_add_file_rmap(page);
2011 set_pte_at(mm, address, page_table, entry);
2013 /* No need to invalidate - it was non-present before */
2014 update_mmu_cache(vma, address, entry);
2015 lazy_mmu_prot_update(entry);
2017 pte_unmap_unlock(page_table, ptl);
2018 return VM_FAULT_MINOR;
2020 page_cache_release(page);
2023 return VM_FAULT_OOM;
2027 * do_no_page() tries to create a new page mapping. It aggressively
2028 * tries to share with existing pages, but makes a separate copy if
2029 * the "write_access" parameter is true in order to avoid the next
2032 * As this is called only for pages that do not currently exist, we
2033 * do not need to flush old virtual caches or the TLB.
2035 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2036 * but allow concurrent faults), and pte mapped but not yet locked.
2037 * We return with mmap_sem still held, but pte unmapped and unlocked.
2039 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2040 unsigned long address, pte_t *page_table, pmd_t *pmd,
2044 struct page *new_page;
2045 struct address_space *mapping = NULL;
2047 unsigned int sequence = 0;
2048 int ret = VM_FAULT_MINOR;
2051 pte_unmap(page_table);
2052 BUG_ON(vma->vm_flags & VM_PFNMAP);
2055 mapping = vma->vm_file->f_mapping;
2056 sequence = mapping->truncate_count;
2057 smp_rmb(); /* serializes i_size against truncate_count */
2060 new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2062 * No smp_rmb is needed here as long as there's a full
2063 * spin_lock/unlock sequence inside the ->nopage callback
2064 * (for the pagecache lookup) that acts as an implicit
2065 * smp_mb() and prevents the i_size read to happen
2066 * after the next truncate_count read.
2069 /* no page was available -- either SIGBUS or OOM */
2070 if (new_page == NOPAGE_SIGBUS)
2071 return VM_FAULT_SIGBUS;
2072 if (new_page == NOPAGE_OOM)
2073 return VM_FAULT_OOM;
2076 * Should we do an early C-O-W break?
2078 if (write_access && !(vma->vm_flags & VM_SHARED)) {
2081 if (unlikely(anon_vma_prepare(vma)))
2083 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2086 copy_user_highpage(page, new_page, address);
2087 page_cache_release(new_page);
2092 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2094 * For a file-backed vma, someone could have truncated or otherwise
2095 * invalidated this page. If unmap_mapping_range got called,
2096 * retry getting the page.
2098 if (mapping && unlikely(sequence != mapping->truncate_count)) {
2099 pte_unmap_unlock(page_table, ptl);
2100 page_cache_release(new_page);
2102 sequence = mapping->truncate_count;
2108 * This silly early PAGE_DIRTY setting removes a race
2109 * due to the bad i386 page protection. But it's valid
2110 * for other architectures too.
2112 * Note that if write_access is true, we either now have
2113 * an exclusive copy of the page, or this is a shared mapping,
2114 * so we can make it writable and dirty to avoid having to
2115 * handle that later.
2117 /* Only go through if we didn't race with anybody else... */
2118 if (pte_none(*page_table)) {
2119 flush_icache_page(vma, new_page);
2120 entry = mk_pte(new_page, vma->vm_page_prot);
2122 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2123 set_pte_at(mm, address, page_table, entry);
2125 inc_mm_counter(mm, anon_rss);
2126 lru_cache_add_active(new_page);
2127 page_add_anon_rmap(new_page, vma, address);
2129 inc_mm_counter(mm, file_rss);
2130 page_add_file_rmap(new_page);
2133 /* One of our sibling threads was faster, back out. */
2134 page_cache_release(new_page);
2138 /* no need to invalidate: a not-present page shouldn't be cached */
2139 update_mmu_cache(vma, address, entry);
2140 lazy_mmu_prot_update(entry);
2142 pte_unmap_unlock(page_table, ptl);
2145 page_cache_release(new_page);
2146 return VM_FAULT_OOM;
2150 * Fault of a previously existing named mapping. Repopulate the pte
2151 * from the encoded file_pte if possible. This enables swappable
2154 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2155 * but allow concurrent faults), and pte mapped but not yet locked.
2156 * We return with mmap_sem still held, but pte unmapped and unlocked.
2158 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2159 unsigned long address, pte_t *page_table, pmd_t *pmd,
2160 int write_access, pte_t orig_pte)
2165 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2166 return VM_FAULT_MINOR;
2168 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2170 * Page table corrupted: show pte and kill process.
2172 print_bad_pte(vma, orig_pte, address);
2173 return VM_FAULT_OOM;
2175 /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2177 pgoff = pte_to_pgoff(orig_pte);
2178 err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2179 vma->vm_page_prot, pgoff, 0);
2181 return VM_FAULT_OOM;
2183 return VM_FAULT_SIGBUS;
2184 return VM_FAULT_MAJOR;
2188 * These routines also need to handle stuff like marking pages dirty
2189 * and/or accessed for architectures that don't do it in hardware (most
2190 * RISC architectures). The early dirtying is also good on the i386.
2192 * There is also a hook called "update_mmu_cache()" that architectures
2193 * with external mmu caches can use to update those (ie the Sparc or
2194 * PowerPC hashed page tables that act as extended TLBs).
2196 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2197 * but allow concurrent faults), and pte mapped but not yet locked.
2198 * We return with mmap_sem still held, but pte unmapped and unlocked.
2200 static inline int handle_pte_fault(struct mm_struct *mm,
2201 struct vm_area_struct *vma, unsigned long address,
2202 pte_t *pte, pmd_t *pmd, int write_access)
2208 old_entry = entry = *pte;
2209 if (!pte_present(entry)) {
2210 if (pte_none(entry)) {
2211 if (!vma->vm_ops || !vma->vm_ops->nopage)
2212 return do_anonymous_page(mm, vma, address,
2213 pte, pmd, write_access);
2214 return do_no_page(mm, vma, address,
2215 pte, pmd, write_access);
2217 if (pte_file(entry))
2218 return do_file_page(mm, vma, address,
2219 pte, pmd, write_access, entry);
2220 return do_swap_page(mm, vma, address,
2221 pte, pmd, write_access, entry);
2224 ptl = pte_lockptr(mm, pmd);
2226 if (unlikely(!pte_same(*pte, entry)))
2229 if (!pte_write(entry))
2230 return do_wp_page(mm, vma, address,
2231 pte, pmd, ptl, entry);
2232 entry = pte_mkdirty(entry);
2234 entry = pte_mkyoung(entry);
2235 if (!pte_same(old_entry, entry)) {
2236 ptep_set_access_flags(vma, address, pte, entry, write_access);
2237 update_mmu_cache(vma, address, entry);
2238 lazy_mmu_prot_update(entry);
2241 * This is needed only for protection faults but the arch code
2242 * is not yet telling us if this is a protection fault or not.
2243 * This still avoids useless tlb flushes for .text page faults
2247 flush_tlb_page(vma, address);
2250 pte_unmap_unlock(pte, ptl);
2251 return VM_FAULT_MINOR;
2255 * By the time we get here, we already hold the mm semaphore
2257 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2258 unsigned long address, int write_access)
2265 __set_current_state(TASK_RUNNING);
2267 inc_page_state(pgfault);
2269 if (unlikely(is_vm_hugetlb_page(vma)))
2270 return hugetlb_fault(mm, vma, address, write_access);
2272 pgd = pgd_offset(mm, address);
2273 pud = pud_alloc(mm, pgd, address);
2275 return VM_FAULT_OOM;
2276 pmd = pmd_alloc(mm, pud, address);
2278 return VM_FAULT_OOM;
2279 pte = pte_alloc_map(mm, pmd, address);
2281 return VM_FAULT_OOM;
2283 return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2286 #ifndef __PAGETABLE_PUD_FOLDED
2288 * Allocate page upper directory.
2289 * We've already handled the fast-path in-line.
2291 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2293 pud_t *new = pud_alloc_one(mm, address);
2297 spin_lock(&mm->page_table_lock);
2298 if (pgd_present(*pgd)) /* Another has populated it */
2301 pgd_populate(mm, pgd, new);
2302 spin_unlock(&mm->page_table_lock);
2306 /* Workaround for gcc 2.96 */
2307 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2311 #endif /* __PAGETABLE_PUD_FOLDED */
2313 #ifndef __PAGETABLE_PMD_FOLDED
2315 * Allocate page middle directory.
2316 * We've already handled the fast-path in-line.
2318 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2320 pmd_t *new = pmd_alloc_one(mm, address);
2324 spin_lock(&mm->page_table_lock);
2325 #ifndef __ARCH_HAS_4LEVEL_HACK
2326 if (pud_present(*pud)) /* Another has populated it */
2329 pud_populate(mm, pud, new);
2331 if (pgd_present(*pud)) /* Another has populated it */
2334 pgd_populate(mm, pud, new);
2335 #endif /* __ARCH_HAS_4LEVEL_HACK */
2336 spin_unlock(&mm->page_table_lock);
2340 /* Workaround for gcc 2.96 */
2341 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2345 #endif /* __PAGETABLE_PMD_FOLDED */
2347 int make_pages_present(unsigned long addr, unsigned long end)
2349 int ret, len, write;
2350 struct vm_area_struct * vma;
2352 vma = find_vma(current->mm, addr);
2355 write = (vma->vm_flags & VM_WRITE) != 0;
2358 if (end > vma->vm_end)
2360 len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2361 ret = get_user_pages(current, current->mm, addr,
2362 len, write, 0, NULL, NULL);
2365 return ret == len ? 0 : -1;
2369 * Map a vmalloc()-space virtual address to the physical page.
2371 struct page * vmalloc_to_page(void * vmalloc_addr)
2373 unsigned long addr = (unsigned long) vmalloc_addr;
2374 struct page *page = NULL;
2375 pgd_t *pgd = pgd_offset_k(addr);
2380 if (!pgd_none(*pgd)) {
2381 pud = pud_offset(pgd, addr);
2382 if (!pud_none(*pud)) {
2383 pmd = pmd_offset(pud, addr);
2384 if (!pmd_none(*pmd)) {
2385 ptep = pte_offset_map(pmd, addr);
2387 if (pte_present(pte))
2388 page = pte_page(pte);
2396 EXPORT_SYMBOL(vmalloc_to_page);
2399 * Map a vmalloc()-space virtual address to the physical page frame number.
2401 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2403 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2406 EXPORT_SYMBOL(vmalloc_to_pfn);
2408 #if !defined(__HAVE_ARCH_GATE_AREA)
2410 #if defined(AT_SYSINFO_EHDR)
2411 static struct vm_area_struct gate_vma;
2413 static int __init gate_vma_init(void)
2415 gate_vma.vm_mm = NULL;
2416 gate_vma.vm_start = FIXADDR_USER_START;
2417 gate_vma.vm_end = FIXADDR_USER_END;
2418 gate_vma.vm_page_prot = PAGE_READONLY;
2419 gate_vma.vm_flags = 0;
2422 __initcall(gate_vma_init);
2425 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2427 #ifdef AT_SYSINFO_EHDR
2434 int in_gate_area_no_task(unsigned long addr)
2436 #ifdef AT_SYSINFO_EHDR
2437 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2443 #endif /* __HAVE_ARCH_GATE_AREA */